Cursor: Cheap Models Write Code to a Frontier Model's Plan, Cutting Cost 15x

In July 2026, Cursor ran an experiment: it had a swarm of AI agents rewrite the SQLite database engine in Rust, armed only with documentation, with no source code, no internet access and no ready-made tests. The new version of the agent swarm, which separates a planner model from worker models, ultimately passed 100% of the test suite in every configuration, and the cheapest solution cost 15 times less than a single top-tier agent.

What Exactly Cursor Tested

Cursor gave the agents a single task: recreate SQLite in Rust from scratch, working only from the official documentation. The finished result was checked against the sqllogictest benchmark, which contains millions of SQL queries. According to Cursor, every configuration of the new system ultimately reached 100% of tests, whereas the old swarm got stuck in the 11–77% range.

The key idea of the new architecture is a split in roles. An expensive frontier model acts as the planner: it breaks the task down and makes architectural decisions. Cheap worker models write the bulk of the routine code.

• The task: rewrite SQLite in Rust from documentation only, with no source code and no internet
• Planners: GPT-5.5, Grok 4.5, Opus 4.8, Fable 5
• Worker: Composer 2.5 (also run as a single agent for comparison)
• Verification: the sqllogictest benchmark, millions of SQL queries
• The new system after 4 hours: 73–85% at the start, all reached 100%; the old one: 11–77%

Why the Planner Removes the Chaos

Splitting the roles frees the swarm from the merge conflicts it used to create for itself. In the Grok 4.5 configuration, the old system made 68,000 commits in two hours and generated more than 70,000 merge conflicts, then choked on them. The new swarm kept up a pace of about 1000 commits per second and never exceeded 1000 simultaneous conflicts.

The difference also shows in the amount of code at comparable quality. The "Fable 5 plans + Composer writes" pairing came in at 9908 lines versus 64,305 for the old approach. The "Opus 4.8 + Composer" pair produced 4645 lines versus 19,013. Less code means fewer bugs and cheaper maintenance.

How Much Was Saved

The cost of the runs differed by 15x at comparable results: from $1339 for the "Opus 4.8 + Composer" pairing to $10,565 for a single GPT-5.5. According to Cursor, the worker models consumed 69–90%+ of all tokens, but because of the price gap their contribution to the final bill was far more modest. Most of the expensive spending goes to planning.

The frontier model takes on decomposition and the key decisions, while cheaper worker models execute the plan efficiently once the ambiguity has been removed.

That is the Cursor team's main conclusion from the experiment.

What This Means

The economics of AI coding are shifting: it makes sense to pay for a top model at the planning stage, while the implementation itself can be handed to cheap workers with almost no loss in quality. For teams, this means a multiple-fold reduction in the cost of autonomous code generation, provided someone smart slices the task correctly.

Frequently Asked Questions

What is an agent swarm in Cursor?

An agent swarm is a group of many AI agents working in parallel on a single task. In the new version, Cursor split them into a planner (one expensive model) and workers (many cheap ones), which eliminated mass merge conflicts.

Which models planned, and which wrote the code?

The planners were GPT-5.5, Grok 4.5, Opus 4.8 and Fable 5. The code was written by Composer 2.5, which was also tested on its own as a single agent for comparison.

How much cheaper was the code?

The spread was 15x: from $1339 for the Opus 4.8 + Composer pairing to $10,565 for a single GPT-5.5, at comparable result quality.